Index Smoothing And Autoregressive Fusion Forecasting Model And Its Empirical Study

The research of time series prediction theory and model has been a hot issue of many scholars,which has important application value.The classical time series prediction model has a solid theoretical foundation and has achieved remarkable results in various fields.The neural network model based on machine learning is highly intelligent.In the era of big data,it has quickly become the preferred model in various fields.However,when one-step prediction is carried out for the time series with non-stationary and less historical data,the applicability of the existing prediction model is not high.Therefore,this paper proposes a one-step prediction method for non-stationary time series with less historical data.Firstly,the existing prediction methods are systematically sorted out,comprehensively studied and focused on.From the two dimensions of method characteristics and historical context,the time series analysis methods are summarized according to the prediction models based on mean feature,difference feature,artificial neural network,chaotic feature and NLP framework.Secondly,aiming at the one-step prediction problem of non-stationary time series with less historical data,the idea of exponential smoothing and autoregressive prediction is combined,and a prediction model based on differential dynamic equation is proposed.The model includes the second-order differential equation describing the prediction mechanism,the discretization structure of the equation,and the network topology.The application principle of the model is given.Two evaluation indexes,prediction error and prediction homomorphism are introduced,and two evaluation indexes,steady-state index and transition index are defined.Finally,based on 12 groups of sample data series,the prediction effectiveness of the model is empirically analyzed.The results show that the absolute value series of prediction deviation of12 groups of samples approximately follows exponential distribution,so the occurrence of large prediction error events is controllable.The model in this paper is suitable for one-step prediction of time series.There is a certain relationship between steady-state index and prediction error,turning index and prediction homomorphism.The prediction error and prediction homomorphism can be estimated by the priors of steady-state index and turning index,which makes the model have a certain priori perception of prediction effectiveness.In view of the systematic deviation of strong trend time series prediction,the proposed algorithm is effective.Compared with LSTM and GRU models based on literature data,the results show that the proposed model is more suitable for one-step prediction than LSTM and GRU models in literature.In addition,the prediction principle of the model is extended to the whole period prediction of periodic time series.Compared with the literature method,the paper model has the advantages of small amount of historical data,dynamic update and use of data,and fast response speed of the model,but the prediction error is slightly higher than that of the literature method.Therefore,as a one-step prediction model,at present,the paper model can only be used as a reference for long-term prediction.This paper has 29 figures,8 tables and 89 references.